标签：#ai

--- title : " AI Harness: the worst and the best buzzword in the industry" published : false tags : [ ai , harness , middleware , finops , aws , bedrock , opensource ] series : " TokenOps on AWS" cover_image : # TODO: circuit-breaker / middleware diagram --- AI Harness: the worst and the best buzzword in the industry "El mercado habla de 'AI Harness' como si fuera magia. El verdadero arnés de un LLM es un Proxy Inverso y un Middleware Transaccional determinístico. Es el código tradicional (styrr-llm y sayay-guard) el que confina, audita y presupuesta la inferencia probabilística antes de que toque tu infraestructura en la nube." — TokenOps raw research, Turno 8 The Hook "Harness" is the most polarizing word in AI engineering right now. Depending on who you ask it's either the industry's worst buzzword or the best technical concept ever packaged badly. It's both — and the difference is whether you can name the actual engineering underneath. Why It's the WORST Buzzword (the smoke) It's a wrapper. 90% of the time, "we built an Enterprise AI Harness" means someone wrote a Python requests script or an Express server that wraps the OpenAI or Bedrock API. Language appropriation. "Harness" literally means arnés — a tether. Marketing sells it as "an intelligent structural armor that tames the wild energy of AI." In systems engineering it's a middleware, or a glorified try/catch with JSON schema validation. No standard. No rigorous CS definition exists, so anyone calls anything "harness" — a log interceptor, a proxy, a YAML config file — inflating expectations without delivering real value. Why It's the BEST Buzzword (the engineering) Strip the LinkedIn marketing and the original test harness metaphor becomes genuinely powerful for generative AI: electrical isolation of uncertainty. An LLM is a highly unstable, probabilistic component. You cannot wire it directly into a bank's production database. You need a physical code "harness" that isolates it. When the model goes crazy

I Built an Autonomous AI Agent That Hunts Bounties. Here's What Happened. The Setup I gave an AI agent one job: find paid work online, build the deliverable, and earn money — autonomously. Not a chatbot. Not a copilot. An agent that scans 232+ listings across multiple platforms, filters out scams and ghost sponsors, writes proposals, generates deliverables with real market data, and queues everything for human approval. Here's what happened in the first 48 hours. The Stack (All Free) Python core — pipeline orchestration, economic gate, critic Ollama + qwen3:4b — local LLM for analysis writing (no API costs) Chart.js — dashboard visualizations Public APIs — CoinGecko, DeFiLlama, Solana RPC (all keyless) GitHub Pages — free hosting for the portfolio Windows Task Scheduler — runs every day at 9 AM + every 4 hours Total infrastructure cost: $0/month. What the Agent Actually Does Every Morning 09:00 — Wake up ├── Check-in on AgentHansa (earn $0.01 USDC daily drip) ├── Scan Superteam Earn (232 live listings) ├── Scan Clawlancer/TaskForce/MoltJobs for gigs ├── Scan GitHub for paid issues ($20-500 fixes) ├── Filter through 7 anti-scam layers: │ geo restrictions, human-presence demands, │ ghost sponsors (no web/twitter/verification), │ unverified payers, real-money requirements ├── Economic gate: expected value must be positive ├── Local LLM critic reviews against actual page content └── If candidate passes everything: → Build deliverable (report/dashboard/thread draft) → Generate proposal text → Send Telegram alert with approval command The Filters That Saved Me In the first 24 hours, the agent found 232 listings. After filtering: Filter Killed HUMAN_ONLY access 216 Ghost sponsors (no identity) 1 (would've wasted hours) Real-money deposit required 1 ($1000 bug bounty trap) Country walls 1 (Superteam Canada only) Already claimed/stale Rest Without these filters, I would have wasted days on bounties that were never going to pay. The First Deliverable The agent found a $500 bo

According to Xbox CEO Asha Sharma, Project Helix, which she announced in March as a codename for Microsoft's "next generation console" - phrasing that seemingly implied a singular device - will actually be a "family" of devices." "We've been hard at work on a great next generation and a great family of devices for Helix, […]

Waymo has launched Gemini in Waymo , a beta in-car conversational assistant for riders using its Ojai robotaxi experience . Accessed through a Gemini icon on the cabin screen, the feature lets riders use natural language and voice to adjust parts of the cabin, ask about their journey and get information about nearby places or broader topics. The important boundary is clear: Gemini is a rider-facing assistant, not part of the autonomous driving system. Waymo Driver continues to control the vehicle , while Gemini operates separately and does not influence driving decisions. For riders, the integration turns the cabin display into a more conversational interface. For the wider automotive market, it is a concrete example of generative AI being deployed inside a commercial mobility service without being assigned responsibility for vehicle control. Waymo describes the feature in its official Gemini in Waymo announcement . The company says Gemini stays inactive until a rider chooses to engage it. It does not access real-time driving data unless the rider explicitly asks for information related to the ride. What Gemini in Waymo can do today Gemini in Waymo is designed around requests that are useful during a trip, rather than around autonomous navigation. A rider can tap or press the Gemini icon and speak to the assistant. The initial beta supports interactions such as: Cabin-control requests , including asking to set the air conditioning to a specified temperature. Ride-related questions , such as seeking information about the current journey. Information about surroundings , including questions about local sites. General knowledge queries through a hands-free conversational interface . This scope matters because it places Gemini in the passenger experience layer. The assistant can make a ride feel more responsive without creating confusion about which system is responsible for safety-critical driving functions. Area Gemini in Waymo Waymo Driver Primary role Rider-facing c

Google has introduced a dedicated Student Hub in the Gemini ecosystem , bringing study notebooks, flashcards and interactive practice quizzes into one in-app space. The central idea is to connect a learner's course materials with Gemini's AI tools, reducing the work of moving between separate note-taking, revision and question-generation tools. The official Gemini for Students page presents the hub as a gateway to Gemini's education-focused capabilities. It is part of a broader Google education AI initiative that also involves NotebookLM and Google for Education resources, rather than a standalone feature with no connection to the rest of Google's products. For students, the practical value is straightforward: uploaded learning materials can become organized revision assets. For businesses that create internal training or support education programs, the release is also a useful example of how generative AI can consolidate material preparation, knowledge review and self-assessment into a more connected workflow. Google has not, however, confirmed a specific learning management system integration in the supplied materials. How Gemini Student Hub connects learning materials and AI tools The Student Hub is designed as a dedicated space where courses and content connect with Gemini. Its core tools include a study notebook, flashcard creation and quick practice quizzes. Google says Gemini notebooks can take uploaded course materials, including PDFs, slides and notes, and generate study aids such as flashcards, quizzes and study guides. A significant detail is the use of inline citations to user-provided sources for those generated materials. That does not remove the need for learners to check the results, but it gives them a way to trace an AI-produced prompt or explanation back to the material they uploaded. In a learning workflow, that is more useful than treating a general-purpose chatbot response as an unanchored answer. NotebookLM is an important part of the wider wo

Google is offering eligible college students in the United States 12 months of Google AI Pro at no charge . The offer, announced on August 19, 2026, gives students access to the paid Gemini plan that Google values at $19.99 per month. It is redeemable through December 31, 2026, and standard Google AI Pro pricing applies after the free year unless the student cancels. The program is aimed at academic work, but it also matters for the wider Gemini ecosystem . It puts higher-capacity AI tools, Google app integrations and substantial cloud storage in the hands of students who may carry those workflows into internships, startups and future workplaces. For businesses, the immediate lesson is not that Google has announced a broader pricing reduction. It has not. Rather, teams should expect more new users to become familiar with Gemini and the ways it connects with everyday Google tools. What Google AI Pro includes for eligible US students According to Google's official student offer announcement , eligible US college students who claim the promotion receive one year of Google AI Pro. Google says the plan includes four times higher usage limits within Gemini , Gemini Spark, integrations with Google apps such as Gmail and Docs , and 5 TB of Google One storage. Google has also introduced a student hub in the Gemini app for participating students. The hub is intended to support learning with features including study notebooks and Deep Research in Gemini Live. These tools are presented as part of a student-focused experience, rather than as a separate business plan or a new API offering. The distinction matters. Access to Gemini through this offer does not, by itself, establish access to every Google AI product or developer service. Students and organizations considering Gemini for a particular workflow should check the relevant product terms and capabilities rather than assuming that an app subscription covers all Google AI services. Offer detail Eligible college students in t

Nobody likes phone trees. "Press 1 for billing, press 2 for support." Miss an option? Start over. It is friction at its worst. The voice-ivr-with-agent-backend example replaces that with a natural language conversation. Callers just say what they need, and the app routes them to the right department. Code: https://github.com/team-telnyx/telnyx-code-examples/tree/main/voice-ivr-with-agent-backend What it builds A Python/Flask app that handles inbound calls with a conversational IVR: Inbound Call -> answer with Call Control -> look up menu config from KV -> LLM generates a dynamic greeting -> gather(speech) — caller says what they need -> LLM routes intent to a department -> transfer call The core primitives The app combines four Telnyx primitives: Call Control : answer() , speak() , gather_using_speech() , transfer() AI Inference : telnyx.ai.openai.chat.completions.create() for greetings and intent routing KV store : menu config per phone number (business name, departments, transfer numbers, keywords) Agent state machine : an IVRAgent class that tracks call state, turn count, and retry logic Dynamic greeting via LLM Instead of a hardcoded "Press 1 for billing," the app generates a conversational greeting from the KV config: def generate_dynamic_menu_prompt ( menu_config : dict ) -> str : departments = menu_config . get ( " departments " , []) dept_list = " \n " . join ( f " - { d [ ' name ' ] } : { d [ ' description ' ] } " for d in departments ) return ( f " You are an IVR assistant for { menu_config [ ' business_name ' ] } . " f " Available departments: \n { dept_list } \n\n " f " Greet the caller briefly and ask how you can help. " f " Keep it conversational and under 2 sentences. " ) The LLM generates the greeting through the OpenAI-compatible Telnyx Inference binding. If it fails, the app falls back to a static greeting from the KV config. Intent routing via LLM When the caller speaks, the transcription is passed to route_intent_with_llm . The LLM is instructed

AI coding agents are getting very good at writing code. They can build components, create APIs, fix bugs, and implement features from short prompts. But I kept noticing one issue: Working code does not always mean a good product. For example, if you ask an agent: “Add a delete button to every project.” It may technically do exactly that. But will it also think about: confirmation before deletion error handling undo options accessibility clear feedback to the user Those are not just coding problems. They are product judgment problems. That led me to experiment with a reusable instruction layer for AI coding agents at AudranLab. The idea is simple: Instead of only asking an agent, “Can you build this?”, also encourage it to ask, “Is this a good way to build it?” I want agents to consider things like accessibility, failure states, destructive actions, usability, and sensible defaults while they work. This does not magically turn an AI into a product designer. But I think it raises an interesting question: Can explicit product principles consistently improve the quality of software generated by coding agents? That is what I’m currently exploring. My next step is to test the approach across different coding tasks and compare the results with and without the additional product-judgment layer. If you’re interested in AI agents, LLM reliability, developer tools, or applied AI, I’ll be sharing more experiments here. AudranLab: https://www.audrantechlab.online/

Neocloud Lambda has raised $1B in private debt to buy Nvidia AI chips and lease them to Microsoft. It's the latest in a string of loans, underscoring the high cost of the AI boom.

Given 10 benchmarks for specific misaligned behaviors, the automated systems were able to improve performance on every single one without degrading overall performance.

Anthropic’s recent work on Claude Sonnet 5 points to a potentially important direction in AI safety: using post-training methods to improve the behavior of increasingly capable models. Public material from Anthropic indicates that Sonnet 5 received substantial post-training alignment work and delivered safety improvements over earlier Sonnet versions. A separate public signal suggests researchers may be exploring whether one model can help align a stronger successor, although the specific reported training lineage has not been documented in Anthropic’s first-party materials. For businesses deploying advanced AI, the practical lesson is not that alignment has been solved. It is that model behavior can be materially shaped after base training, and that safety results need to be assessed in the context of the tasks a company actually plans to automate. What Anthropic’s published results establish In its official Claude Sonnet 5 announcement , Anthropic describes substantial post-training intended to align the model with Claude’s constitution. The company reports improvements in safety-related behavior, including stronger refusals of unsafe requests and lower misalignment findings in automated audits compared with Sonnet 4.6. That is meaningful because post-training is the stage where a model’s responses, instruction-following behavior, and safety boundaries can be adjusted after its underlying capabilities are developed. In operational terms, it can affect whether an AI assistant follows risky instructions, mishandles sensitive workflows, or produces responses that conflict with a company’s intended rules. However, the available research also establishes an important limit. Sonnet 5 was not uniformly at the level of Claude Opus 4.8 across every safety measure. Anthropic’s evaluations still identified some automated assessments where Sonnet 5 showed higher misalignment relative to Opus 4.8. Opus 4.8, released in May 2026, is the company’s production-ready reference poin

Un déploiement devrait être la chose la plus ennuyeuse de ta semaine. S'il est excitant, c'est mauvais signe. Au début de ma carrière, les mises en production étaient des événements. On retenait son souffle, on croisait les doigts, quelqu'un exécutait de mémoire une séquence d'étapes manuelles, et on regardait les journaux avec une boule au ventre. C'était palpitant. C'était aussi terrifiant, et le côté palpitant était précisément le problème : chaque déploiement était un pari, parce que chaque déploiement était un peu différent du précédent. Un bon déploiement est répétable. La même chose, de la même façon, à chaque fois — automatisée, pas récitée par un humain fatigué à la fin d'une longue journée. Quand le processus est un script plutôt qu'une cérémonie, l'ennui remplace l'angoisse. Tu ne pries plus. Tu appuies sur un bouton, et le résultat est prévisible parce qu'il a déjà été prévisible cent fois. L'automatisation fait ici plus que gagner du temps. Elle supprime toute une catégorie d'erreurs : l'étape oubliée, le mauvais paramètre, le « je croyais que tu l'avais fait ». La machine ne se fatigue pas, ne saute pas de ligne, ne se laisse pas distraire à mi-chemin. Elle rend le déploiement fiable au point d'en être ennuyeux — et l'ennui, en production, est un luxe. Alors, si tes mises en production font encore monter le rythme cardiaque, ce n'est pas de la prudence. C'est un signal. Rends-les répétables, rends-les automatiques, rends-les ennuyeuses. Garde le frisson pour ta vie ; ton système de production, lui, mérite l'ennui. – Serguey Shinder

Anthropic’s publicly documented work on AI safety includes Petri , an open-source behavioral auditing tool, and ongoing updates to Claude models such as Claude Opus 4.7 . Those materials show continued investment in testing model behavior and improving model capabilities. They do not, however, substantiate a precise claim that Claude improved safety scores across 10 alignment failures without capability trade-offs, or that particular methods generalized to models exactly 4.7 times larger. That distinction matters for teams evaluating AI systems. Broad statements about alignment progress can be useful signals of research direction, but operational decisions need to rest on documented evaluations, relevant use cases, and the controls a company can apply in its own workflow. Anthropic’s public record supports a narrower, more practical conclusion: behavioral auditing is becoming a more visible part of how frontier AI models are assessed, while model releases and safety research remain separate evidence streams. What Anthropic’s public materials document Petri is designed for behavioral AI auditing Anthropic describes Petri as an open-source auditing tool . Its Petri 2.0 update, published in January 2026, added a larger seed library with 70 new seeds and improved mitigations intended to address evaluation awareness. Evaluation awareness is relevant because a model may behave differently when it appears to be taking a test than when it is operating in a more ordinary setting. The Petri 2.0 work reported results across 10 target models , using Claude Sonnet 4.5 and GPT-5.1 as auditors. This establishes that Anthropic has described a cross-model auditing effort. It does not establish that Claude itself achieved a safety improvement across 10 defined alignment failures. A target-model count, an auditor model, and a set of alignment failures are different measurements and should not be treated as interchangeable. For readers, the important point is that behavioral audits can

Le maillon le plus faible de ta sécurité a un pouls. Ce n'est pas ton pare-feu, ni ton chiffrement, ni ton dernier correctif. C'est une personne — et les attaquants le savent bien mieux que la plupart des équipes. Pourquoi forcer une porte blindée quand on peut simplement demander la clé ? La majorité des intrusions sérieuses ne commencent pas par un exploit technique génial. Elles commencent par un e-mail qui a l'air juste assez vrai, un appel qui semble venir du service informatique, une pièce jointe qu'une personne pressée ouvre sans réfléchir. La technologie tient. C'est l'humain qu'on contourne. Cela dérange, parce que c'est plus difficile à corriger qu'une faille logicielle. On ne corrige pas les gens. Mais on peut les préparer. La formation ne consiste pas à traiter les employés d'imprudents ; elle consiste à leur montrer à quoi ressemble vraiment une attaque, pour qu'ils la reconnaissent dans un moment de fatigue. Et il faut concevoir en supposant que quelqu'un se fera avoir un jour. Parce que quelqu'un se fera avoir. L'authentification à plusieurs facteurs, le moindre privilège, la limitation de ce qu'un compte compromis peut atteindre : tout cela existe précisément parce qu'un humain finira par cliquer sur le mauvais lien. La question n'est pas si, mais quand — et ce qui reste debout après. Alors ne consacre pas tout ton budget aux murs et rien aux personnes qui gardent les portes. Le maillon le plus faible a un pouls, un mauvais jour, et une boîte de réception pleine. Protège-le comme le reste de ton infrastructure, parce que c'en est la partie la plus exposée. – Serguey Shinder

TL;DR: Running high-frequency autonomous AI agent loops on commercial LLM APIs at scale is economically unsustainable and introduces unpredictable latency spikes. This production guide details how we deployed a self-hosted inference cluster using vLLM (v0.6+) , EAGLE-3 speculative decoding , PagedAttention v2 , and Automatic Prefix Caching (APC) on cloud GPUs (RunPod/Vast.ai), achieving a sub-180ms Time-To-First-Token (TTFT) , 118 tokens/sec throughput , and cutting inference costs by 45–74% . 1. The Economic & Latency Bottleneck of Agentic Loops When building 24/7 autonomous daemon agents , LangGraph multi-agent state machines , or LLM-driven NPC game loops , the computational profile differs fundamentally from human chatbot interactions: Massive Request Volume: A single complex agent decision cycle frequently executes 5 to 25 LLM calls across intent classification, tool schema validation, reflection loops, and output formatting. Repeated Prefix Redundancy: 80–90% of prompt tokens consist of identical system instructions, persona framing, and MCP (Model Context Protocol) tool definitions. Strict Latency Budgets: Real-time simulations and game loops cannot tolerate 800ms–1500ms commercial API network roundtrips. Commercial Closed APIs (GPT-4o / Claude 3.5 Sonnet) ├── Prefill: Paid per-token on every single cyclic call ├── Network Roundtrip: 250ms - 600ms latency overhead └── Cost at 50,000 daily agent iterations: $1,200 - $3,500 / month Self-Hosted vLLM Cluster (RTX 4090 / A100 on RunPod) ├── Automatic Prefix Caching (APC): Reuses KV-cache (120ms -> 12ms prefill) ├── Speculative Decoding (EAGLE-3): 2.1x generation throughput └── Fixed Infrastructure Cost: $245 - $480 / month (Flat, unlimited tokens) 2. Deep Dive: The vLLM Memory & Scheduling Architecture PagedAttention: Eliminating KV-Cache Fragmentation Standard PyTorch/HuggingFace transformer implementations allocate static KV-cache tensors sized for max_sequence_length . Because 95% of queries generate far fewer

On ne gère pas ce qu'on ne mesure pas. C'est l'une des premières leçons de l'exploitation, et pourtant je l'ai apprise à l'envers, en pilotant à l'aveugle bien trop longtemps. Sans mesure, tu ne sais pas si un système va bien. Tu le supposes. Il tourne, personne ne se plaint, donc tout va bien — jusqu'au jour où quelque chose se dégrade lentement, sous le radar, et où tu ne l'apprends que lorsque c'est déjà une panne. La lente fuite de mémoire, le disque qui se remplit, la latence qui grimpe d'une milliseconde par semaine : rien de tout cela ne crie. Ça glisse. La mesure transforme les suppositions en faits. Un tableau de bord, quelques alertes bien choisies, et soudain tu vois le problème arriver au lieu de le subir. Tu n'attends plus que l'utilisateur t'apprenne que ton système est cassé ; tu le sais avant lui. Mais il y a un piège que j'ai appris à éviter : mesurer trop. Cent métriques que personne ne regarde ne valent pas mieux que zéro. Le bruit noie le signal, et les alertes qui se déclenchent sans raison finissent par être ignorées — jusqu'à celle qui comptait vraiment. Bien mesurer, ce n'est pas tout mesurer. C'est choisir les quelques signaux qui prédisent réellement un problème. Alors, avant de bâtir la prochaine chose, demande-toi comment tu sauras si elle va mal. Si la réponse est « quelqu'un finira par le remarquer », tu ne la gères pas encore. Tu espères. Et l'espoir n'est pas une stratégie d'exploitation. – Serguey Shinder

Alle paar Jahre verkündet jemand, dass die IT-Jobs verschwinden. Die Cloud ersetzt die Administratoren. Die Automatisierung ersetzt die Operatoren. Jetzt ersetzt die KI, was übrig ist. Und alle paar Jahre verschwinden die Jobs nicht, sie wandern. Sie wandern den Stack hinauf. Als physische Server der Cloud wichen, brauchten wir keine Infrastruktur-Leute weniger, wir brauchten Menschen, die Infrastruktur auf einer höheren Ebene verstehen, die Systeme entwerfen, absichern und kostenbewusst betreiben, statt Hardware in Racks zu schrauben. Als manuelle Deployments der Automatisierung wichen, brauchten wir Menschen, die die Automatisierung selbst bauen konnten. Jede Welle entfernte nicht die Arbeit. Sie hob den Boden und verschob den wertvollen Teil nach oben. Die KI ist die nächste Welle, und ich erwarte, dass sie sich genauso verhält. Sie wird viel Routine aufsaugen: das Skripten, die ersten Config-Entwürfe, die Standard-Fehlersuche. Was sie nicht aufsaugt, ist Urteilsvermögen: zu wissen, was zu bauen ist, zu entscheiden, was das Risiko wert ist, zu verstehen, wie die Teile einer echten Organisation zusammenpassen, und geradezustehen, wenn etwas schiefgeht. Wer in jedem Übergang strauchelt, sind die, die sich über die Aufgabe definieren, die automatisiert wurde. Wer gedeiht, definiert sich über das Problem, das er löst, und lässt die Werkzeuge dafür sich darunter ändern. Verteidige also nicht das eine, das du heute tust. Werde gut in der Schicht darüber. Der IT-Job hat jedes Werkzeug überlebt, das ihn beenden sollte, indem er den Stack hinaufwanderte. Diesen wird er genauso überleben. – Serguey Shinder

There's a lot of capital pouring into the business of giving models away.

James Cameron's 1991 sci-fi blockbuster returns to theaters this weekend for its 35th anniversary.

Just as new data centers face growing backlash from neighboring communities, the US Environmental Protection Agency (EPA) is about to make it harder for people to weigh in on any pollution those centers create. The EPA plans to toss out a federal rule requiring public notice and an opportunity to comment when certain industrial sites […]